Centralization's Impact on Mutual Fund Performance: Decentralized vs. Centralized Funds, Lecture notes of Decision Making

Download Centralization's Impact on Mutual Fund Performance: Decentralized vs. Centralized Funds and more Lecture notes Decision Making in PDF only on Docsity!

Does Firm Organization Matter? Evidence from

Centralized and Decentralized Mutual Funds

MARCIN KACPERCZYK

New York University and NBER

AMIT SERU

University of Chicago and NBER

March, 2012

Does Firm Organization Matter? Evidence from

Centralized and Decentralized Mutual Funds

Abstract

We examine the impact of centralization of investment decisions by a fund’s family on its funds’ performance and show that funds from decentralized families have higher performance than their centralized counterparts. We exploit a quasi-experiment involving failed mergers to gen- erate exogenous variation in acquisition outcomes of target funds. A difference-in-differences estimation reveals that, relative to failed funds, those acquired in a merger by centralized (de- centralized) fund family produce lower (higher) performance. These differences in performance are driven by more discretion in managerial decision making in decentralized fund families. We confirm these findings by tracing the response of centralized and decentralized funds to exoge- nous changes in their information environment. Though funds in centralized families have lower performance than their decentralized counterparts, we show that these families allow for better coordination in trading and brokerage decisions and better diversification across funds in the family.

information. Other forms of incentive compensation may not be able to fully solve this problem since a large part of this information may be information that may be difficult to transfer (soft information) to committees or other decision-making bodies (Stein [2002]) or may be non-contractible (Aghion and Tirole [1997]). In contrast, the downside of a decentralized decision-making organization is the lack of coordination in decisions across the agents in the firm.

These economic trade-offs generate several testable predictions. First, we expect decen- tralized mutual funds to perform better than their centralized counterparts. This should be largely on account of decentralized funds using more private information in their investment decisions. Second, since the manager talent pool is not homogeneous, we expect higher ability managers to get attracted to decentralized funds—where the decision making allows for better use of the more valuable information they can produce. The first part of the paper uses panel data on nearly 3000 actively managed U.S. equity funds to test these predictions and finds strong support.

The caveat with the empirical approach in the first part of the paper is that omitted factors could impact both fund family decision to follow a certain decision making and their investment and performance. We exploit a quasi-experiment that helps circumvent this challenge. The empirical design uses failed mergers of funds to generate exogenous variation in acquisition outcomes of target funds. In particular, we construct a group of target firms whose mergers failed to go through for reasons that are unrelated to investment skill of the target (‘control group’). We then assemble a ‘treatment group’ comprising of funds taken over in a merger. The two groups then comprise a sample where we claim that the assignment of a firm into a fund family is orthogonal to investment performance (or its expectation) at the time of merger. Under this assumption, we can difference out any selection bias by comparing the investment performance of the funds in the treatment group pre and post-merger with those of the control group.

A difference-in-differences-in-differences estimation reveals that, relative to failed funds, those acquired in a merger by centralized fund family produce lower performance relative to funds in the control group. Strikingly, funds involved in a merger where the acquirer has a decentralized structure see an improvement in their performance. This experiment lends support to the notion that differences in organizational decision making are responsible for generating differences in fund performance.

We next investigate the mechanism that helps generate the differences in performance. We start by showing that there are large differences in the extent of private information that is generated by the two fund structures – with more private information being produced in de-

centralized fund families. We find this inference holds in the quasi-experiment involving failed mergers as well in the overall pooled sample. We also confirm these findings by tracing the response of the two fund structures to exogenous variation in information quality about invest- ment opportunities faced by them. We find that an exogenous reduction in quality of public information about investment opportunities leads to a differential increase in performance of decentralized mutual funds relative to centralized ones. This evidence is also consistent with centralized funds producing less private information—and as a result suffering more relative to decentralized funds when the precision of public information falls after the shock.

We further extend our analysis by providing direct evidence on the mechanism that drives differences in performance across two structures. In particular, we show that decentralized funds offer greater discretion to their managers, measured by the number of funds per manager. Moreover, they also provide higher-powered incentives, in that fund managers in those families are more likely to be promoted (demoted) based on their superior (inferior) performance.

Overall, we provide strong evidence that decentralized fund families perform significantly better and are likely to produce more private information. In equilibrium, such a result would be hard to justify as it appears that decentralized funds offer greater benefits to their investors. We argue that the downside of a decentralized decision-making organization is the lack of coordination in decisions across the agents in the firm.

Empirically, we demonstrate this to be case. Funds in centralized families allow for better coordination in trading and brokerage decisions across funds in the family. In particular, centralized families are less likely to cannibalize each others’ trades in that they are less likely to trade in the opposite directions at the same time. Moreover, funds in centralized families are more likely to use similar execution brokers which is likely to lower their execution costs.

We also provide evidence that centralized fund structures provide better diversification opportunity for their investors. The reason for this, again, involves better coordination of investment decisions that results in smaller “diversification loss” (see, Sharpe [1981]). In summary, there is some evidence that centralized fund structures allow for better coordination of investment decisions across funds in the family.

The empirical literature on the impact of organization design on investment decisions is small. For instance, Mullainathan and Scharfstein [2001], Baker and Hubbard [2004], and Ciliberto [2006] investigate the implications of organizational form on various investment de- cisions. Similarly, Beshears [2010] and Seru [2010] assess if organizational form impacts firm’s productivity. However, most of these studies are indirect since they do not directly observe inputs and outputs of agents undertaking decisions. In contrast, our study extends this liter- ature by providing us a setting where we can observe both the impact on inputs (information)

families in the world, Fidelity Investment Management. In its official statement to shareholders, Fidelity states:

“At Fidelity, individual portfolio managers are ultimately responsible for investment decisions. Since our founding, we have believed that individual responsibility and accountability for investment decisions is much more effective than decisions made by committee.”

This would constitute a decentralized decision making fund family. In contrast, T.Rowe Price could constitute a centralized decision making fund family since it manages portfolios using a team approach where members across funds work closely together on the investment strategy, asset allocation, portfolio construction, and security selection.^1 Given the theoretical literature discussed above, we expect mutual fund managers who are allowed more flexibility in deciding on their investment strategies to produce more valuable information. The testable prediction that emerges from this discussion is that, ceteris paribus, funds that are part of decentralized investment management companies should produce more private information in their investment decisions and as a result outperform their centralized counterparts.

The downside of a decentralized organization, however, is the lack of coordination in deci- sions across the agents in the firm. In other words, we expect better coordination across funds in the family which has centralized decision making in the family. This idea is exemplified in the following report from Barclays Global Management:

“Barclays Global Investors offers performance management through quantitative, structured and index management techniques. This focus eliminates the traditional active management decision process and provides maximum cost control.”

The testable prediction that emerges from this discussion is that, ceteris paribus, central- ized investment management companies should exhibit better coordination across investment decisions in their funds. In our analysis, we focus on two measurable dimensions. First, we examine the proportion of trades in opposing directions across funds in a given family during a time period. In a similar vein, we also investigate the number of brokers handling the trades across funds in a family during a time period.

Second, we also assess if centralized investment management companies may allow for bet- ter diversification of risk across its funds for investors. There could be several reasons for (^1) It is instructive to note that both structures seem to manage significant amounts of assets on behalf of their clients. In 2009, in a $12-trillion mutual fund industry, about 40% of fund families had centralized structures while the remaining had 60% decentralized structure.

diversification loss in decentralized fund family. First, in a decentralized fund family, invest- ment decision in each fund is likely to have more influence of individual fund manager. As a result, there may be manager-specific risks which are introduced across funds in a decen- tralized fund family that are not as easily diversified. Second, it is also possible that multiple individual managers in a decentralized fund may not account for the correlation of their own portfolio returns with the returns of other managers in the fund family. This diversification loss in decentralized fund family is similar to the notion discussed in Sharpe [1981] and later advanced by Elton and Gruber [2004] and van Binsbergen, Brandt, and Koijen [2008].^2

We end this section by noting that the differences in type of incentives offered in the two organizational forms may induce managerial sorting if managers are relatively informed about their ability. Specifically, if higher-ability managers prefer discretion in their investment decision making, they might sort into a decentralized fund family. In the empirical analysis, any differences in performance across different organizational structures could be driven by differences in managerial quality. To the extent that this heterogeneity in managerial quality is induced by differences in discretion across organizations, this self-selection does not pose threat to our empirical analysis. However, there may be other reasons for heterogeneity in managerial quality across the two organizational forms. Therefore, our empirical identification is devoted to providing evidence for the causal link between organizational form and fund performance after conditioning for managerial quality.

III Data Description

Our main data set is the Center for Research on Security Prices (CRSP) survivorship bias- free mutual fund database. Our full sample covers the period 1980-2005. Given the nature of our tests and data availability, we focus on actively managed open-end U.S. equity mutual funds. However, for our measure of diversification and some tests we utilize the data of all fund families covered by CRSP. We further merge the CRSP data with fund holdings data from Thomson Financial. The total number of funds in our merged sample is 3477.

We also use the CRSP/Compustat stock-level database, which is a source of information on individual stock returns, market capitalizations, book-to-market ratios, momentum, and standardized unexpected earnings (SUE). In addition, for some of our exercises, we map funds to the names of their managers using information from CRSP, Morningstar, Nelson’ Directory of Investment Managers, Zoominfo, and Zabasearch. This mapping results in a sample with (^2) The diversification loss results from the standard portfolio optimization result that the unconstrained so- lution to the mean-variance optimization problem for the entire portfolio (fund family) as a whole is usually different from the optimal linear combination of mean-variance efficient portfolios (individual funds).

The summary statistics of the data for the performance regressions (time-series averages of cross-sectional means, medians, and standard deviations) are reported in Table I. In a sample of over 190, 000 fund-time observations 32.6% are represented by funds in centralized organization structures. The standard deviation of Central is 46.8%. The annualized average fund return in a sample equals 7.74%. An average fund in our sample exhibits positive stock-picking ability as evidenced by CS value of 0.79% per year; at the same time, consistent with prior literature, an average fund underperforms relative to its passive benchmark, both using 3-factor and 4-factor abnormal returns, with respective values of − 0 .04% and − 0 .17% per year.

Our tests involving coordination across investment decisions across funds in a family will include the following family-specific control variables: The family size (natural logarithm of total net assets under management in million of dollars obtained as a sum of all funds inside the family, Log(Famsize)), the family age (natural logarithm of family age in years since in- ception, Log(Famage)), the family return (value-weighted return of all funds inside the family, Famreturn), number of funds inside the family (natural logarithm of the number, Log(Funds)), and percentage value of equity funds inside the family (% Equity).

IV Preliminary Evidence

We begin our analysis by assessing differences in performance of funds that belong to decen- tralized fund families relative to those that belong to centralized ones. To do so, we estimate a regression model of the following form:

Performanceit+1 = α + βCentralit + γ′Xit + μt + μi + it+1, (1)

where Performance is a generic variable for four different performance measures: Return, CS, 3-Factor α, and, 4-Factor α. The right-hand side includes lagged value of Central as an explanatory variable with β measuring the impact of organization’s decision making on performance. We saturate the empirical specification with a plethora of explanatory variables that might be important in explaining the variation in fund performance. In particular, in vector X, we include Log(Assets), Log(Age), Turnover, Expense, Load, and style attributes (Size), value (Value), and momentum (Momentum). We additionally include fund and time fixed effects, denoted by μ.^4 Since Central varies at the family level, we cluster standard errors at the fund family groupings. (^4) Note that, in the specifications that use fund fixed effects, the identification on the coefficient of Central in this section is coming from changes in decision making faced by some funds in the data. Over the sample period, we have 140 changes where the investment decision making of a fund changes. This is largely on account of fund mergers changing the decision making associated with its family.

The regression results are presented in Table II. We first present results for specifications without time fixed effects (columns (1)-(3)). In column (1), we present results for CS. The coefficient of Central equals − 0 .88%, a drop of 4% of a standard deviation, and is statistically significant at the 1% level. In column (2), for 3-factor α, the coefficient of Central is − 0 .74%, a drop of 10% of a standard deviation, and is statistically significant at the 5% level. For 4-factor α, in column (3), the effect is 0.56% reduction, an equivalent of 7% of standard deviation, and is statistically significant at the 10% level. We include time-fixed effects in columns (4)-(6) and find that the results are quantitatively very similar. These results provide evidence that is consistent with the notion that funds in centralized structures have lower performance relative to those in decentralized structures.

V Identification

The evidence from panel data suggests that organizational structure plays an important role in performance differences across funds. However, such evidence is difficult to interpret as causal, especially since the context of our analysis makes it subject to potential endogeneity concerns. We start by describing a selection consideration in this spirit—managerial sorting— which may make such inferences difficult. We provide evidence for such selection and offer a simple solution to deal with this concern. We then discuss results from a quasi-experiment that more generally accounts for selection bias due to omitted variables (including managerial sorting).

V.A Differences in Managerial Quality

The simplest concern with interpreting our performance results causally is that these differences could arise from several other factors. Not all such factors are necessarily inconsistent with our arguments on differences in organization of the the two fund structures being responsible for differences in performance. However, the presence of these factors makes it hard to isolate the effect of fund organization on performance, all else equal. The simplest example concerns the heterogeneity in managerial quality in the two structures. Why might we expect quality of managers to vary depending on the fund structure? Under our interpretation, decentralized structures provide managers with more flexibility to execute their ideas. As a result, we expect managers with better investment skills sort into these type of organizations. There could be other factors too. But regardless of the source of this sorting, its presence makes interpreting the results from Table II difficult.

We first conduct analysis to assess if there are differences in managerial quality across

performance among fund managers not observed to the econometrician, but conditioned upon by outside labor market, we use return gap – the measure of unobserved actions introduced in Kacperczyk, Sialm, and Zheng [2008], Return Gap. We additionally include fund and time fixed effects. Standard errors are clustered at family groupings. The coefficient of interest is β, which measures the relative importance of the structure for the subsequent career outcomes.

We report the results in Panel B of Table III. In column (1), we consider External Promotion and find that managers employed by centralized structures are less likely to be promoted to a larger fund outside her own family. The effect is statistically significant at the 10% level of significance. Similarly, in column (3) we show that such managers are less likely to land a job in a hedge fund industry. This effect is statistically significant at the 1% level. At the same time, while we find, in column (2), that managers in a centralized structure are more likely to be demoted to a smaller fund outside their own families, this effect is statistically insignificant. Overall, this evidence suggests that managers of higher quality sort into decentralized fund families.

V.A.1 Simple solution

The results in Section V.A indicate a significant heterogeneity in managerial ability, possibly induced by differences between the two organization structures. These results, while consistent with our hypothesis, also pose an important challenge on how to interpret our findings. Could the results on differences in performance be driven by differences in managerial ability—with higher quality managers matched to decentralized funds for unobservable reasons other than our hypothesis? To the extent that the differences in performance in the two structures are not driven by time-invariant manager characteristics, introducing manager-fixed effects in our previous specifications should isolate the effect of organization structure.^5 In this section, we consider this possibility.

In our analysis, we present the results from panel data—but rather than estimating the performance regression at the fund level, we reorient our data to the manager level to facilitate the use of manager-fixed effects. In columns (1)-(3) of Table IV, we present the results with fund and time-fixed effects, but without manager-fixed effects. We obtain similar results to those we obtained in our fund-level specification, in Table II. Subsequently, in columns (4)- (6), we estimate the same model adding manager-fixed effects to the specification. Including manager-fixed effects retains the economic and statistical significance of our findings. Note that these specifications are more stringent and exploit the within manager-fund variation. In (^5) Introducing fixed effects accounts for time-invariant but not time-varying managerial attributes. However, the essence of our empirical context indicates that the former variation is likely to be more important in this setting.

other words, the identification here comes from tracking the performance of the same manager across a centralized and decentralized fund structures over time. These results indicate that differences in structure itself maybe an important driver of our findings.

Overall, the results suggest that while differences in managerial ability may be driving some of the magnitudes, a large part of our findings are also influenced by differences in organization structure. While it could still be the case that we may not have accounted for some omitted factor which affects both differences in manager ability and the performance across the two structures, these specifications provide comfort that the time invariant component of such a factor is unlikely driving our findings on performance.

V.B Evidence from a quasi-experiment on failed fund mergers

In this section, we exploit a quasi-experiment using failed fund mergers to handle selection concerns more generally. Before proceeding with the analysis it is worth clarifying the nature of selection issue more formally. The goal of empirical tests so far has been to isolate the causal effect of centralized activity on manager investment behavior. Let us represent this average treatment effect as AT E = E[yi(C = 1) − yi(C = 0)], where yi(C = j) is the investment performance of a fund i when it is (not) a part of centralized family j = 1 (j = 0). To illustrate the complications in making causal inferences, let us focus on regressions in Table II where we observed E[yi(1)|C = 1] − E[yi(0)|C = 0], i.e., the difference in average performance of centralized firms relative to decentralized ones. It is useful to note that:

E[yi(1)|C = 1] − E[yi(0)|C = 0] = (E[yi(1)|C = 1] − E[yi(0)|C = 1])

• ( ︸E [yi(0)|C = 1] (^) ︷︷− E[yi(0)|C = 0])︸

The bracketed term is the ‘selection bias’ that plagues the estimates. Put simply, it says that there might be a bias in the estimates since investment performance of mutual funds that are a part of the centralized firm might differ from those that are decentralized—both on observables and unobservables. If one could randomly assign funds with similar investment skill into centralized and decentralized funds, one could remove this bias (since with random assignment E[yi(0)|C = 1] = E[yi(0)|C = 0]).^6 Following this, the empirical design in this part of the paper is geared towards discussing a quasi-experiment that tries to get as close as possible towards generating this random assignment. (^6) It is easy to show using iterated expectations that the term that survives after the selection bias is removed is equal to ATE.

are very similar in terms of different measures of performance.

In the unreported test we pooled all target firms and examine whether the investment performance of the targets at any time t − 1 can predict the deal’s success at time t. We take all the years till the year in which the deal either succeeds or fails (inclusive). More concretely, the specification is:

P rob (Successit = 1) = Φ

α + (^) γXit− 1 + β 1 Performanceit− 1 + μt

where Φ denotes the logit distribution function and X are control variables. The dependent variable Success takes a value 1 for the treatment group in the event year and 0 otherwise. The coefficient of interest is β 1 , the coefficient of performance of the target fund prior to the merger event.

In this regression, we find that β 1 is insignificant across specifications. This evidence suggests that, conditional on other observables, the pre-merger characteristics of the control and the treatment sample are quite similar on the dimension of performance—and as a result, measures of investment are not able to predict which deal will subsequently succeed. This analysis validates the methodology that was used to construct the control sample.

V.B.2 Main Result

In our main tests, we estimate a difference-in-differences-in-differences (DDD) specification that compares the investment performance of targets within the treatment and control groups before and after the event dates and then compares the difference across the two groups for mergers involving centralized acquirers versus decentralized ones. Specifically, the specification that is estimated in Table V is:

P erf ormanceit =

α + γ 1 Afterit + γ 2 Afterit ∗ Ti + γ 3 Afterit ∗ Ti ∗ Centrali + δZit + μt + μi

where, After is an indicator variable that takes a value 1 for all the years after the event date and 0 otherwise and T is an indicator variable that takes a value 1 for targets in the treatment group and 0 for targets in the control group. And, Central is an indicator variable that takes a value 1 for mergers that involve centralized acquirers and is 0 for mergers where the acquirer is decentralized. All the regressions are estimated with time (μt) and fund (μi) fixed effects.

Note that, this specification allows us to assess the change in performance of funds in the treatment group when the mergers are ones where acquirers have decentralized headquarters relative to the funds in the control group (γ 2 ). In addition, it also helps to infer the change in performance in instances when the funds in the treatment group are acquired by centralized

fund family, relative to the funds in the control group (γ 3 ). The net effect of acquiring targets in a centralized structure would therefore be given by γ 2 + γ 3.

In Panel A, we consider the three-factor alpha as a measure of performance. As can be observed across specifications, γ 2 > 0, suggesting that funds that are acquired by decentralized families tend to improve their performance after the merger, relative to similar funds in the control sample. This implies that the funds that are acquired by families that are decentral- ized are able to leverage on better environment for private information generation (including appropriate incentives and resources)—thereby outperforming funds that do not successfully merge.

Strikingly, we observe an opposite pattern for funds that are acquired by centralized fund families, relative to similar funds in the control sample (γ 3 < 0). This suggests, consistent with our hypothesis, that such fund organizations are not conducive to producing private information that helps in superior performance. Moreover, the net effect of such mergers, given by γ 2 + γ 3 is negative. This suggests that, overall, mergers where the decision-making for acquired fund becomes centralized see a drop in their performance after the merger event. The results described here remain similar across specifications that iteratively adds more controls as well as clusters standard errors at time dimension.

In Panel B, we present the results with alternative measures of performance. The results in the first two columns use four-factor alpha and confirm the inferences made in Panel A. Interestingly, we find no action on the dimension of holdings-based measures. We present results using CT measure, but the nature of the results are similar for other holdings-based measures as well. Overall, the findings from this section confirms that differences in fund organization has a direct effect on the performance of the funds.

VI Economic Mechanism

Our analysis suggests that that decentralized structures produce better performance than their centralized counterparts. In this section, we provide evidence that drills down on the mechanism that drives our findings. In Section VI.A, we provide evidence that managers in decentralized structures outperform their counterparts in centralized structures because they produce more private information. In addition, we also provide evidence that suggests that decentralized structures provide fund managers with more discretion and incentives to produce private information for their investment activities in Section VI.B.

We also test for this effect directly in the quasi-experiment discussed earlier. In particular, we follow the same specification as (4) but use RPI as the dependent variable. We present the results in Panel B of Table V. As is shown in Columns (5) and (6), the results suggest that there is a dramatic drop in reliance on private information subsequent to the merger for cases when the fund is acquired by centralized organization (γ 3 < 0). No such effect is evident when the merger involves a decentralized organization. This result confirms that the mechanism driving differences in performance across the two fund structures is related to to the amount of private information being produced in these organizations.

We also conduct additional tests that confirm this mechanism. In particular, we directly study the resources for information production employed by the fund organizations. To the extent that fund managers of decentralized structures would like to produce more information internally (instead of relying on public sources), the organizations employing them would likely facilitate such process and supply the private information to their trading desks. As a result, we expect more personnel associated with collecting private information on investment opportunities to be present in decentralized structures. This personnel could include buy- side security analysts, security traders, and other fund managers, among others. We gather information on these variables from Nelsons’ Directories. In addition, we use family size, Famsize, as an indirect proxy for the ability of the fund family to provide useful information following the intuition that large companies, such as Fidelity, are more likely to be better equipped to provide useful information to their managers.

Formally, we estimate the following regression model:

InfoProductionit = α + βCentralit + γ′Xit + , (6)

where InfoProduction is a generic variable for any measure of the internal information pro- duction sources. In our regression model, we alternately use the number of analysts, Ana- lysts, the number of managers, Managers, and the number of security traders employed by the family, Traders. The last specification includes the natural logarithm of family assets, Log(Famsize). Our main variables of interest is Central. We also include a vector of control variables, Log(Famage), Famreturn, and Expenses. We cluster standard errors at the family groupings. Our coefficient of interest is β, which measures the differential response of the two organizational structures in terms of their information production. Panel B of Table VI presents the results.

In columns (1) and (2), we respectively consider the univariate and multivariate specifi- cations for Analysts. Consistent with our prediction, we find a strong negative effect of a centralized structure on the number of analysts employed. The coefficient of Central is sta-

tistically significant at the 1% level of significance; it is also economically significant: Moving from centralized to decentralized structure amounts to an increase of about 9 analysts. In columns (3) and (4), we examine the effect of structure on Managers. Like in the previous case, we again find a strong and negative effect of a centralized structure on the number of employed managers. The coefficient of Central is statistically significant at the 1% level of sig- nificance. Moving from a centralized to decentralized structure leads to an increase of about 13 managers, an economically significant result. Finally, in columns (5)-(6), we look at the effect on Traders and find a similarly strong effect: Moving from away from centralized structure amounts to an increase of about 2.5-3 security traders. Taken together, these results indicate that decentralized funds have more information-gathering resources which is consistent with funds in these structures producing more private information.

VI.A.1 Evidence from information environment shock

We now provide additional evidence that supports the notion that decentralized fund structures produce superior performance on account of more private information. In particular, we exploit changes in information environment about investment opportunities faced by different fund structures and trace responses of fund performance to these shocks. Specifically, our empirical design builds on the notion that fund managers weigh public and private information in their investment decisions. Consequently, if managers in decentralized structures generate more precise private information about their investment opportunities, an exogenous change in the precision of public signal about investments should impact the reliance on public information and the investment performance of decentralized funds to a lesser degree.^8

Our experiment requires an exogenous shock to precision of public information of invest- ments across various funds. For that purpose, we build on Hong and Kacperczyk [2010] (HK) and use mergers between large brokerage houses. The sample construction is discussed in Appendix. HK show that such mergers result in the firing of analysts because of redundancy (e.g., one of the two analysts covering the same stock is let go) and other reasons such as culture clash – and this reduction impacts analyst forecast bias. We verify in Appendix that reduction of analysts following mergers of brokerage houses also leads to an increase in analyst (^8) This follows from an idea that was exploited before in Kacperczyk and Seru [2007]. In a setting similar to Grossman and Stiglitz [1980] they show that managers who produce more private information put smaller weight on the public signal. More formally, suppose all participants observe a normally distributed public signal about the asset fundamental value with mean ¯u and precision ρpu. Moreover, suppose informed managers observe a private signal, θ, about the asset fundamental value, u, with θ = u+ , where  is distributed N (0, ρ). Then any informed manager calculates the expected value and precision of the asset mean as, E[u | θ] = (^) ρ+ρρpu θ+ (^) ρρ+puρpu ¯u, with ρ[u | θ] = ρ + ρpu. Thus if the precision of private signal (ρ) is tied to skill (see Kacperczyk and Seru [2007]), we would expect that any exogenous change in the precision of the public signal should impact the holdings, and the investment performance of decentralized funds to a lesser degree.